ANCIENT COINS & CERAMICS - 3D AND 2D DOCUMENTATION FOR PRESERVATION
AND RETRIEVAL OF LOST HERITAGE
Hubert Mara†, Martin Kampel‡, Franco Niccolucci† and Robert Sablatnig‡
† PIN - Servizi Didattici e Scientifici per
l’Universit di Firenze
VAST-Laboratory
Piazza Ciardi 25, 59100 Prato, Italy
hubert.mara@pin.unifi.it, niccolucci@unifi.it
http://www.pin.unifi.it
‡ Vienna University of Technology
Institute for Computer Aided Automation
Pattern Recognition and Image Processing Group
Favoritenstrasse 9/183-2, 1040 Vienna, Austria
kampel@prip.tuwien.ac.at, sab@prip.tuwien.ac.at
http://www.prip.tuwien.ac.at
KEY WORDS: 3D-Acqisition, Cultural Heritage, Ceramics, Coins, Rotational Axis, Symmetry Analysis, Pattern Recognition
ABSTRACT:
Motivated by the requirements of archaeologists we are developing an automated system for acquisition, documentation and management of daily finds of excavations. These daily finds can be separated into large objects like remainders of architecture and small
objects of ancient daily life - like ceramics and coins. Ceramics especially are found in numbers of tens of thousands on virtually every
excavation, because ceramics have been in use for thousands of years. Until the present day these finds are documented by manual
drawings. There is a similar situation in the case of coins, where manual drawings are often used to abstract them from photographs.
Therefore we propose an automated acquisition and documentation system based on digital cameras and structured light for small
findings. For ceramics we provide further processing to estimate horizontal cross-sections (profile-lines) for printed documentation, as
it is done by manual drawing. For this a proper orientation of the acquired 3D-model is required and automatically estimated based on
the assumption that ceramics were made on rotational plates (wheels).
We are aware that ceramics might not always have been manufactured on rotational plates, because the wheel was not invented everywhere as for the example in the Americas. Even though ceramics from such areas appear to be rotational symmetric, we developed a
method based on shape and symmetry analysis to determine the manufacturing techniques of ceramics. This helps to answer another
archaeological question regarding the technological advance of an ancient culture.
Results for accuracy and performance are shown on real data from recent interdisciplinary projects together with archaeologists from
Austria, Germany, Israel and Peru. Furthermore we present preliminary results of the integration of coin classification in our documentation system. Additionally we are currently adapting the London Charter to ensure the intellectual integrity, reliability, transparency,
documentation, standards, sustainability and accessibility of the information gathered by 3D-acquisition. Finally a summary and an
outlook is given.
1
INTRODUCTION
Documentation of ceramics is a main task in archaeology, because ceramics are the most common findings, used and produced
in large numbers by humans for several thousands of years. Archaeologists use analysis of ceramics (Leute, 1987) on a daily
basis to reveal information about the age, trading relations, advancements in technology, art, politics, religion and many other
details of ancient cultures. The same applies to coins, which are
found in small numbers, but they have a similar importance, because they can be related to time periods and regions thanks to
the embossed information e.g. like well- known emperors.
Both types of finds – ceramics and coins – are typically documented by manual drawings and not photographs, because a
draftsperson draws only the relevant details of a find. This has
to be done, because the finds generally have signs of wear, dirt,
corrosion, etc. which tricks the human eye. Furthermore these
finds have to be documented properly, because they cannot be
brought outside the countries hosting the excavation. Also, due
to security issues even local access to the storages of the finds is
granted to limited groups of experts.
Therefore we are developing two systems: the first system is an
automated system for ceramics documentation to help archaeologists document their finds in an efficient and accurate way, which
can be used for further (computerized) research. The second system aims to retrieve finds, which are often stolen from excavations due to their value for collectors. While the first system has
evolved within several national and international projects, the development of the second system has been recently started within
COINS: Combat On-line Illegal Numismatic Sales-project.
At a first glance these two systems appear to be separate, but
they will be united within future work, because the methodologies will be interchanged resulting in a 2D and 3D-framework
for documentation and retrieval of small finds. Therefore it must
be stressed that the methods for retrieval of stolen coins can be
extended and adopted to ceramics to (virtually) retrieve and therefore enable the determination of the context between finds of different excavations.
Therefore the first and main part regards ceramics as described
in Section 2 showing basic documentation for daily finds and
specialized methods, which help archaeologists to answer more
specific questions about ceramics. The second part (Section 3)
shows preliminary results of the COINS-project. Finally a summary and an outlook about documentation and preservation of
small archaeological findings are given in (Section 4).
2
CERAMICS
The basis of documentation of ceramics is a manually drawn horizontal intersection, which is called profile line (Leute, 1987).
As the ceramics are found in tens of thousands at virtually every excavation, these drawings require a lot of time, skill and
man-power of experts. Therefore we are assisting archaeologists
in interdisciplinary projects (Kampel and Sablatnig, 1999, Cosmas et al., 2001) by using an automated system for acquisition
and documentation of ceramics using a 3D-scanner based on the
principle of structured light (DePiero and Trivedi, 1996, Liska,
1999).
This section is structured into three parts: First we describe the
methods used for basic documentation (Section 2.1), which can
be applied on complete ceramics and their fragments. This work
is the basis for symmetry-analysis leading to the determination
of the manufacturing technique of ancient ceramics, as shown in
Section 2.3. Finally Section 2.4 shows an example for processing
decorated ceramics for further archaeological research.
2.1
ACQUISITION
Regarding documentation of ceramics the basic method applied
to unbroken ceramics and fragments of ceramics is the estimation of a profile-line. This is the longest elongation around – or
cross-section through – the wall of a ceramic defined by the rotational axis (also called axis of symmetry). The term rotational
axis relates to the fact that rotational wheels (plates) have been
used for thousands of years for manufacturing ceramics. This assumption can be made especially, but not only for daily finds on
archaeological excavations. Therefore we base our work on using
the rotational axis to orient a ceramic or its fragment to estimate
the profile line as it is done manually by drawings. Furthermore
the creation of such manual drawings is a time-consuming task
requiring expert-skills and our system will help to dramatically
reduce the time for documentation. Figure 11 shows an orientated
front-view and a manual drawing of profile line of a fragment of
a ceramic (also called sherd) found at the excavation at Tel Dor,
Israel (Stern, 1994, extended 2000).
As photography has already proven its reliability for archaeology,
we chose to use a camera and a light-source for 3D-acquisition.
Our first prototype (Sablatnig et al., 1991) in our laboratory was
based on the principle of structured light (DePiero and Trivedi,
1996). Its predecessor was a digital still camera with a special flashlight called Eyetronics ShapeSnatcher (Cosmas et al.,
2001). This camera was developed and used in the 3D-MURALE
project (Cosmas et al., 2001).
Nowadays we use 3D-scanners from the Konica-Minolta Vivid (Mara,
2003, Mara and Hecht, 2006) product range, because of their resolution (≈ 0.1 mm), which meets the requirements given by
archaeologists for their documentation. Figure 2 shows the setup
of our 3D-scanner from recent experiments at the excavations in
the Valley of Palpa (Reindel and Cuadrado, 2001), Peru.
Furthermore Figure 2(a) shows the triangulation principle (Mara,
2003) using a laser (bottom) and a camera (top) having a wellknown distance and orientation. Additionally the turntable – also
shown in this Figure – is used to get a complete 3D-model of the
ceramic. The number of 3D-scans depends on the complexity of
the ceramic and it typically ranges from two scans for sherds up to
eight scans for vessels. The 3D-scans are registered using (Tosovic, 2002) to reassembe a complete 3D-model.
(a)
(a)
(b)
Figure 2: Konica-Minolta Vi-9i 3D-scanner (a) projecting a laserplane (bottom arrow) on to a sherd, while the camera (top) acquires the projection of the laser. Having a well-known distance
and orientation of the laser-plane and the field of view of the camera the distance (range) can be estimated. (b) Detail of this setup
showing a sherd mounted with plasticine on the turntable, which
is used for controlled acquisition of all sides of an object.
2.2
PROCESSING
Therefore the challenging tasks for developing a documentation
system for archaeology are to build a system, which is accurate,
portable, inexpensive, easy-to-use and robust for all kinds of foreign climate, which can range from desert, to jungle to arctic.
This means several technologies like computer tomography and
other laboratory equipment is often not suitable for the daily work
of archaeologists - especially not for ceramics.
After the registration, noise from dust and other objects like holding devices (e.g. clamps or plasticine) are removed from the
3D-model. Then the orientation is estimated based on the assumption that ceramics are rotationally symmetric objects (Mara,
2003), because they were generally manufactured on rotational
plates. The principle of our orientation method is fitting of circle
templates (Gander et al., 1994). In comparison to the computerized, but manual method of (Melero et al., 2003) our orientation method can be used fully- and semi-automatic (Lettner et al.,
2006). Furthermore our system is capable to store the 3D-model
and further archaeological information (e.g. description, photographs, etc.) in a database. For solving the puzzling problems
of other – typically industrially manufactured – rotational objects,
methods like (Pottmann and Randrup., 1998, Willis, 2004, Orriols, 2004) can be applied. Having an orientated 3D-model a
vertical cross-section is estimated using the point of maximum
height of the 3D-model. this cross-section is the so-called profile
line, which concludes the traditional archaeological documentation.
1 The manual drawing was made available by Ilan Sharon, Hebrew
University of Jerusalem.
For comparison with the well consolidated manual approach we
acquired 25 sherds given by archaeologists. All these sherds are
(b)
Figure 1: (a) Front-view and manual drawing1 of a(b) profileline of a sherd found at the excavations at Tel Dor, Israel.
considered ”small”, which means that any sherd smaller than the
selected ones, cannot be orientated manually. For further crossvalidation the sherds have also been orientated and documented
through the use of a Profilograph (Utili and Dolmazon, 2002).
For comparison of the profile lines from these three ways of documentation, we used the novel method of Thodberg (Thodberg,
2003) for shape matching to estimate the differences between
them. This has to be done, because for real data there is no ground
truth in general.
Figure 3 shows one estimated profile lines using manual orientation, the Profilograph and orientation using circle templates.
Further results can are shown in (Mara, 2003). Concluding these
results, we could orientate 24 out of the 25 sherds. Comparison
of the profile lines between the systems showed, that the systems agree for ≈ 2/3 of these sherds, while the remaining 1/3
of the profile lines have identical shape, but differing radii and
orientation in respect to the rotational axis. The reason is that
these sherds are small fragments of large and/or flat vessels (e.g.
plates≈cones).
(a)
(b)
surface with a high resolution (0.1 mm). Therefore we can analyze the symmetry and estimate features like deviation of real
surfaces in respect to a perfectly symmetrical surface. Such features can help archaeologists to decide about the technological
advancements of ancient cultures.
To begin our investigation and answer questions about the manufacturing process of ceramics, we chose to use two modern pots,
which were manufactured in a traditional way. Therefore this data
can be interpreted as mixture between synthetic and real data, because we used real objects. However, unlike real archaeological
fragments, we know how they was produced.
Furthermore we decided to use the method for finding the orientation of a sherd (Mara and Kampel, 2003). We began with the
profile line, which can be estimated in a similar way like in case
of sherds. The difference is that for complete vessels the bottom
plane can be used for orientation, because it is the counterpart for
the rotational plate, which defines the (orthogonal) axis of rotation.
We estimated multiple profile lines, which can be overlaid by
transforming them into the same coordinate system, where the
y-axis equals the rotational axis. Therefore the distance between
profile lines can be estimated. Figure 4 shows the longest profile line and multiple profile lines combined with the side-view,
as archaeologists show such vessels in their documentation. In
the case of the multiple profile lines, we have estimated that the
distance between the profile lines differs and therefore these pots
and their profile lines are unique. The maximum distance between two profile lines of the first pot was 9.8 mm, whereas for
the second pot was 21.2 mm.
Figure 3: (a) Profile lines estimated by manual orientation
(M N ), the Profilograph (P R) and using circle templates (3D).
The dotted profile lines are rotated and (b) shifted by the Thodberg’s method for comparison.
As such a rather simple two-dimensional profile line does not reflect any information about the manufacturing quality leading to
the manufacturing technique, we decided to enhance our system
by giving the archaeologists a tool to gather further information
about the acquired 3D-model. These enhancements are shown in
the following sections. Figure
2.3
(a)
(b)
(c)
(d)
SYMMETRY ANALYSIS
As archaeologists are also excavating burial places, where unbroken ceramics or complete sets of sherds are found, we are
presenting a method to determine the manufacturing process of
ceramics, which reveals information about the technological advancement of an ancient culture. Furthermore this method can be
applied, but is not limited to, unbroken or reconstructed vessels.
Technological advancement, is determined by archaeologists between ceramics, that have been produced either on slow or fast
turning rotational plates. Another example is an ongoing discussion between archaeologists about the existence of rotational
plates for manufacture of ceramics in South America. The general opinion is that in this region the wheel was not invented,
therefore ceramics were produced without a rotational plate (wheel)
(Wieczorek and Tellenbach, 2002) on the other hand-side there is
evidence that rotational plates were used (Carmichael, 1986).
As we use structured light as 3D-acquisition method, we cannot
make an assumptions about the internal structure of a ceramic
as (Wieczorek and Tellenbach, 2002), but we can estimate the
Figure 4: (a,c) Longest profile lines and (b,d) multiple profile
lines of modern ceramics, manufactured in traditional way, which
are supposed to be identical.
In the multiple profile lines shown Figure 4(c,d), the distance between profile lines, measured parallel to the x-axis, is not equal.
If the profile lines were parallel, this would mean, that the pots
have an elliptic (horizontal) cross-section. As it appears, the
asymmetry is more complex. Therefore, we chose to analyze the
pots slice-by-slice along the rotational axis, presumed as orthogonal to the bottom plane.
Figure 5(a,c) shows horizontal intersections, which have been applied with a distance of 10 mm along the rotational axis. The
distance of 10 mm corresponds to the manufacturing process,
which has left traces in the form of rills as seen along the right
hand sides of Figure 4(b,d). These rills are spaced 10 mm, which
corresponds to the width of the finger or tool used to ”grow” the
pot along the axis of the rotational plate. The intersections at
160 mm and 170 mm in height have been discarded, as they intersect the ”shoulder” of the pot with a very low angle (<< 5°),
resulting in an intersection having a non-representative, random
curvature.
Dividing ceramics into sections by characteristic points (like the
”shoulder”) is carried out by archaeologists for classification. Therefore we chose to analyze the object segmented into a lower and
an upper part. This means, we have two fragments, where axis
estimation can be applied, as for sherds (fragments). The estimation of the axis is shown in Figure 5(b,d). The numeric results for
the axis are that they have a minimum distance of 4 mm towards
each other and to the axis defined by the bottom plane. Furthermore the angles between the axes differ between 5° to 7°.
(a)
(a)
(b)
(c)
(d)
(e)
(f)
(g)
(h)
(b)
Figure 6: Axis of rotation and multiple profile lines of the upper
part (a,e), lower part (c,g) and (b,d,f,h) the longest profile lines of
the parts of the objects.
(c)
(d)
Figure 5: (a,c) Top-view and (b,d) side-view of the horizontal
cross-sections - the level of gray corresponds to the height. The
axis of rotation for the lower and upper part is shown as a black
line, defined by the centers of the concentric circles (shown as
dots).
Using the rotational axis of the lower and upper fragment, we repeated the estimation of the profile lines, which are shown in Figure 6. The maximum distance between the profile line are 7 mm
for the upper and 2 mm for the lower part. Therefore the first
conclusion is that the upper and lower parts do have a different
axis of rotation, which means that these parts have been produced
separately and combined without the use of the rotational plate.
We can conclude that, based on the different deviation of the multiple profile lines shown in Figure 6, the upper part is of lesser
quality than the lower part. This leads to the conclusion that these
parts have been made by potters with different experience and/or
on a slower rotational plate. Vice-versa the deviation in the upper part of up to 7 mm compared to less than 2 mm of the lower
part, shows that a faster turning rotational plate has been used that
more experience was required for manufacturing the upper part.
From the differing angle between the axis of rotation based on
the bottom plane compared to the axis of rotation of the upper
and lower fragment, we can conclude that either the bottom has
been post-worked or the pot was contorted before being fired in
the oven.
Even correcting the axis for the parts of the object, the horizontal
intersections are not perfectly circular. The horizontal intersections are elliptic. Therefore we estimated the direction of the
major and minor axis of the ellipses. We estimated that the minor
axis has the same direction as the orientation of the handle. This
means that the symmetry of the pots was broken, when the handle
was attached and the pots were still wet. Figure 7 show the pots,
intersected by a plane, defined by the center of gravity of the pot
and the direction of the major axis of the ellipses. The angle between the minor axis and the handle of the pot was 7° and 14° for
the second pot.
We additionally conclude that the ellipses fitted (Gander et al.,
1994) to the horizontal cross-sections can be used as an additional
feature. Therefore the distance between the foci of the ellipse
is estimated. Ceramics with a distance converging towards zero
(circular cross-sections) are of higher quality.
The proposed method has also been tested on 17 real vessels (Mara
and Abidi, 2001). This approximation is iteratively continued
as ”marching front” (Novotni and Klein, 2002) until no more
vertices with a geodesic distance lower than dist are found. A
schema of the geodesic distance is depicted in Figure 8(a). Figure 8(c) shows a detailed version of one randomly chosen vertex
vi and its geodesic neighborhood. The grey level in Figure 8 corresponds to the geometric distances, darker means lower geodesic
distance.
Step 2: The convolution: The distance of a vertex vi to bi is
denoted as dij . The 1D-filter kernel is defined as:
(a)
Figure 7:
object.
(b)
Planes of symmetry of the (a) first and (b) second
and Hecht, 2006), dated from the NASCA-period (Carmichael,
1986), which were found in the Valley of Palpa, Peru (Reindel
and Cuadrado, 2001). Therefore we could seperate these vessels into three classes determined by the symmetry. The vessels
(60%) of two of these three classes were not produced on rotational plates. Beside this information – answer to the question of
the manufacturing technique – about the use of rotational plates
in South America, this classification is used by archaeologists of
the German Institute for Archaeology (DAI, Bonn) for refinement
of their classification schemes.
2.4
LINE DETECTION ON SURFACES
The approach shown in this section is an example to attempting to
add the analysis of color information of ceramics, which already
have been acquired for automated profile estimation as shown in
the previous sections. The only requirement is to utilize a 3Dscanner capable of texture-mapping. The goal of this example is
to apply image processing methods directly to this type of surface
(texture mapped) images in order to analyze the drawings painted
on surfaces of the pottery.
There are several approaches in the literature that deal with feature detection on surfaces. (Rössl et al., 2000) applies morphological operators to extract feature lines on triangulated surfaces.
A similar approach was presented by (Pottmann et al., 2004) who
apply morphological operators on surfaces, using new metric,
called isophotic metric. While the main goal of these approaches
is to discover characteristics in the curvature of the surface (e.g.
defects on surfaces), we are interested in the texture, i.e. the pictorial information represented by the texture.
A method to enhance linear structures in 2D-images is the convolution of an image with an isotropic second derivative filter
(LoG) (Marr and Hildreth, 1979). An isotropic filter has the advantage, that no directional information from the input data is
necessary. On the other hand, no directional information is provided by the output. In our case, the adaptation for 3D-data is
straight forward and is accomplished in two steps:
Step 1: Definition of the ”convolution window” The image on
the surface is given by a mesh of vertices vi and each vertex is
assigned a grey value gi . For every vertex vi a set of neighbored
vertices bi , i.e. the convolution window, is estimated. This is
achieved by estimation of the distances along the surface (called
geodesic distances) to direct neighbors of vi . The set of vertices of the direct neighbors of vi is called 1-ring. For 1-rings
the geometric distance equals the geodesic distances. So, based
on the geodesic distances of the 1-ring, the geodesic distances
to vi for all neighbors of the 1-ring can be approximated (Sun
f (x) =
1
σ2
−x2
x2
−
1
e 2σ2
2
σ
The discrete convolution is implemented as the sum of a componentwise multiplication of the grey values gj assigned to all neighboredP
vertices bi and vi with the corresponding value of f (di ),
ci = j∈bi ,vi gj f (dij ).
The acquisition system provides two types of data, 3D-range data
and 2D-images of the surface. Both data sets can be related to
each other, e.g. each vertex of the 3D mesh can be assigned
a color value and vice-versa. For our experiments we used the
intensity value, i.e. the I component in the HSI color space,
calculated from the color texture value. For the application of
the derivative Gaussian filter we select all neighboring vertices
with a geodesic distance less than 1.5× the width of the expected
line. This is an empirical value, which was obtained from our
experiments with 2D-line images, where the size of the filter kernel had to be 3 times the width of the lines. Figure 8 shows the
comparison between (c) the manual drawing, (d) the automated
unwrapping (do Carmo, 1976, Strubecker, 1969) of the texturemap of the 3D-model and (e) the detected lines highlighted in
black. Regarding performance this experiment shows that the automated estimation of a drawing similar to the manual drawing
can be obtained within seconds instead of hours. Future work
will add pattern recognition methods (e.g. shape-matching) to
obtain missing parts of the paintings and to automatically find
similar decorations within a collection of 3D-models.
3
COINS
As we do not want to limit our system to ceramics, we recently
started research for integrating the documentation of ancient coins,
another common find, which is typically acquired by digital cameras. Besides automated coin recognition and classification based
on 2D-images, we also want to acquire 3D-data, because medieval coins are not as flat as their modern counterparts. This
would lead to new perspectives of studying ancient coins as well
as new strategies for representing them. Based on the images of
coins for documentation purposes, we propose a pattern recognition system for automatic estimation of an abstract description
and for determination of the type an ancient coin belongs to.
While this can be done quite quickly by an expert, we were asked
by them to apply this system on images found on the web using
web-crawling techniques, because coins are often stolen from excavations and sold on the black market using the internet.
Recent research approaches for coin classification algorithms focus solely on the recognition of modern coins. Applied pattern
recognition algorithms are manifold ranging from neural networks
(M.Fukumi et al., 1992, Bremananth et al., 2005) to eigenspaces
(Huber et al., 2005), decision trees (Davidsson, 1996), edge detection and gradient directions (Nölle et al., 2003, Reisert et al.,
(c)
(a)
Figure 9: Example images of the ancient data set.1
(d)
(b)
(e)
Figure 8: (a) Graphic illustrating the active vertex v1 (black) and
the neighbored vertices b1 within the geodesic distance dist. (b)
Detailed sample of a randomly selected vertex v and its geodesic
neighborhood. The level of grey corresponds to the geodesic distance - darker means closer. Unwrapping of the decoration of
vessel 2801-V3 of the excavation in the Valley of Palpa, Peru:
(c) manual drawing, (d) unwrapped texture and (e) texture with
detected edges (highlighted in black).
2006), and contour and texture features (van der Maaten and
Postma, 2006). (Huber et al., 2005) presents a multistage classifier based on eigenspaces that is able to discriminate between
hundreds of coin classes. Due to the controlled setup of the system presented coin detection becomes a trivial task. They report
correct classification for 92.23% of all 11, 949 coins in a sample set. (van der Maaten and Poon, 2006) presents a coin classification system based on edge-based statistical features, called
COIN-O-MATIC. The system is subdivided into five stages: segmentation, feature extraction, preselection, 3- nearest neighbor
approach, and verification. At the MUSCLE CIS Coin Competition (Nölle et al., 2006) the method achieved a recognition rate of
67.31% on a benchmark set of 10, 000 coins. The coin classification method proposed by (Reisert et al., 2006) is based on gradient information. Similar to the work of (Nölle et al., 2003) coins
are classified by registering and comparing the coin with a preselected subset of all reference coins. The proposed method won
the MUSCLE CIS Coin Competition 2006 (Nölle et al., 2006)
with a recognition rate of 97.24% on a benchmark set of 10, 000
coins.
Tests performed on image collections both of medieval and modern coins show that algorithms performing well on modern coins
do not necessarily meet the requirements for classification of medieval ones (van der Maaten and Postma, 2006). The main difference between ancient and modern coins is that the ancient coins
have no rotational symmetry and consequently their diameter is
unknown. Since ancient coins are all too often in very poor conditions, common recognition algorithms can easily fail. The features that most influence the quality of the recognition process
are yet unexplored. An example of an ancient coin from a dataset
consisting of 3, 000 high resolution images of ancient coins on
constant, white background is shown in Figure 92 . The coins picture Roman emperors from ≈30 B.C. to ≈300 A.D. who form
the 106 classes of the dataset.
In the first stage of our recently initiated research, we started by
analyzing and evaluating the existing algorithms for coin recogni2 The
data set was made available by Klaus Vondrovec, Numismatic
Commission, Vienna, Austria
tion. Inital tests were performed using a MUSCLE CIS database
of coins consisting of over 9, 100 training images (distributed
over 100 classes of coins) and 1, 100 test images. We applied
the algorithm provided by Maaten (van der Maaten and Postma,
2006) based on edge-based statistical distributions. Unlike Maaten,
we completely ignored sensor measures (such as thickness or diameter) and thus we only rely on the extracted features. Our experiments based on the multi-scale edge angle-distance distribution achieved a recognition rate of 63%. Surprisingly, coin classification based solely on area measurement achieved a rate of
45%. Ancient coins differ strongly from modern ones. Crucial
influence has the nature of the medieval coins, which are less details, no rotational symmetry and poor conditions due to abrasion
or fouling. Hence, coin classification based solely on edge-based
features would most probably fail with the classification of ancient coins. In the next stage, we will perform tests on a medieval
database to find those features (or set of features) that most influence the quality of medieval coin representations.
4
SUMMARY AND OUTLOOK
Summarizing the present work shows how our heritage can be
retrieved, when it has been lost or stolen and how it can be preserved using digital documentation using 3D- and 2D-image processing and pattern recognition methods. Furthermore these automated methods help experts of cultural heritage to dramatically
reduce time for documentation of the majority of finds within archaeology and to reduce the time to search numismatic databases
containing thousands of coins.
Beside improvement of existing methods for 3D-Vision another
important work has recently begun to ensure the intellectual integrity, reliability, transparency, documentation, standards, sustainability and accessibility of the information gathered by the
increasing use of 3D-scanners. Therefore we are adopting The
London Charter (Beacham et al., 2006), which will be a future
standard for the use of 3D-Vision within Cultural Heritage.
For future work we will combine the methodologies for coins and
ceramics to achieve a more generic system to retrieve stolen ceramics and to document and manage coins within digital libraries.
Furthermore the methods used to retrieve stolen finds will also
be used to reveal new knowledge hidden within the tremendous
amounts of finds using features of their surfaces.
ACKNOWLEDGEMENTS
The authors would like to thank Paul Kammerer, Vienna University of Technology, Institute for Automation, PRIP Group for
assistance with the adoption of the methods of line-detection to
3D-models.
Furthermore we would like to thank all the archaeological institutes recently involved in our work for providing objects, new
challenges and feedback. These archaeological institutes are:
The (German) Archaeological Institute (DAI, Bonn), the Austrian
Government Agency for Cultural Heritage (Bundesdenkmalamt),
the Museum for History of Arts (Kunsthistorisches Museum) in
Vienna and the Austrian Academy of Sciences (ÖAW).
Finally we would like to thank the European Commission for the
support of the presented work done within the CHIRON-fellowship
(MEST-CT-2004-514539) and the COINS-project (FP6-SSP5-044
450).
Liska, C., 1999.
Das Adaptive Lichtschnittverfahren zur
Oberflächenkonstruktion mittels Laserlicht. Master’s thesis, Vienna University of Technology, Vienna University of Technology, Institute of Computer Aided Automation, Pattern Recognition and Image Processing Group.
Mara, H., 2003. Documentation of rotationally symmetrical archaeological finds by 3d shape estimation. Technical Report
PRIP-TR-103, Vienna University of Technology, Inst. of Computer Aided Automation, Pattern Recognition and Image Processing Group.
REFERENCES
Beacham, R., Denard, H. and Niccolucci, F., 2006. An introduction to the london charter. In: M. I. et al. (ed.), The
e-volution of Information Communication Technology in Cultural Heritage: where hi-tech touches the past: risks and challenges for the 21st century, Short papers from the joint event
CIPA/VAST/EG/EuroMed, Archaeolingua, Budapest, Hungary.
Bremananth, R., Balaji, B., Sankari, M. and Chitra, A., 2005. A
new approach to coin recognition using neural pattern analysis.
In: Proc. of IEEE Indicon 2005 Conference, pp. 366–370.
Carmichael, P., 1986. NASCA Pottery Construction. In: Nawpa
Pacha, Vol. 24, Berkeley, California.
Cosmas, J., Itagaki, T., Green, D., Grabczewski, E., Gool, L. V.,
Zalesny, A., Vanrintel, D., Leberl, F., Grabner, M., Schindler, K.,
Karner, K., Gervautz, M., Hynst, S., Waelkens, M., Pollefeys,
M., DeGeest, R., Sablatnig, R. and Kampel, M., 2001. 3D MURALE: A Multimedia System for Archaeology. In: Proceedings
of the International Conference on Virtual Reality, Archaeology
and Cultural Heritage, Athens, Greece, pp. 297–305.
Davidsson, P., 1996. Coin classification using a novel technique
for learning characteristic decision trees by controlling the degree
of generalization. In: Proc. of 9th Int. Conference on Industrial
& Engineering Applications of Artificial Intelligence & Expert
Systems (IEA/AIE-96), pp. 403–412.
DePiero, F. W. and Trivedi, M. M., 1996. 3-D Computer Vision
Using Structured light: Design, Calibration, and Implementation
Issues. Advances in Computers 43, pp. 243–278.
do Carmo, M. P., 1976. Differential Geometry of Curves and
Surfaces. Prentice-Hall.
Mara, H. and Hecht, N., 2006. 3D-Acquisition and Analysis of
freehand manufactured NASCA Ceramics. In: Proc. of CAA06:
Computer Aplications an Quantitative Methods in Archaeology,
p. to appear.
Mara, H. and Kampel, M., 2003. Automated Extraction of Profiles from 3D Models of Archaeological Fragments. In: O. Altan
(ed.), Proc. of CIPA2003: XIX CIPA Int. Symposium: New Perspectives to Save Cultural Heritage, CIPA 2003 organising committe, pp. 87–93.
Marr, D. and Hildreth, E., 1979. Theory of edge detection. In:
Proc. R. Soc. Lond., B 204, pp. 301–328.
Melero, F., Leon, A., Contreras, F. and Torres, J., 2003. A new
system for interactive vessel reconstruction and drawing. In: Proceedings of CAA’03: Computer Applications in Archaeology,
pp. 8–12.
M.Fukumi, S.Omatu, F.Takeda and T.Kosaka, 1992. Rotationinvariant neural pattern recognition system with application to
coin recognition. IEEE Transactions on Neural Networks 3(2),
pp. 272–279.
Nölle, M., Penz, H., Rubik, M., Mayer, K. J., Holländer, I. and
Granec, R., 2003. Dagobert – a new coin recognition and sorting
system. In: Proc. of the 7th International Conference on Digital
Image Computing - Techniques and Applications (DICTA’03),
Macquarie University, Sydney, Australia., CSIRO Publishing,
pp. 329–338.
Nölle, M., Rubik, M. and Hanbury, A., 2006. Results of the muscle cis coin competition 2006. In: Proc. of the Muscle CIS Coin
Competition Workshop, Berlin, Germany, pp. 1–5.
Gander, W., Golub, G. and Strebel, R., 1994. Least-squares fitting
of circles and ellipses. BIT 34, pp. 558–578.
Novotni, M. and Klein, R., 2002. Computing geodesic distances
on triangular meshes. In: In proc. of the 10th Intl. Conference in
Central Europe on Computer Graphics, Visualization and Computer Vision 2002 (WSCG02).
Huber, R., Ramoser, H., Mayer, K., Penz, H. and Rubik, M.,
2005. Classification of coins using an eigenspace approach. Pattern Recognition Letters 26(1), pp. 61–75.
Orriols, X., 2004. Generative Models for Video Analysis and 3D
Range Data Applications. Ph.D. thesis. Universitat Autonoma de
Barcelona, Spain.
Kampel, M. and Sablatnig, R., 1999. On 3d Modelling of Archaeological Sherds. In: Proceedings of the International Workshop on Synthetic-Natural Hybrid Coding and Three Dimensional Imaging, pp. 95–98.
Pottmann, H. and Randrup., T., 1998. Rotational and helical
surface approximation for reverse engineering. Computing 60,
pp. 307–322.
Lettner, M., Mara, H., Müller, A., Sablatnig, R., Singer, M. and
Krenn, M., 2006. PAT: Profile Analysis Tool for the Documentation of Archaeological Finds. In: Proc. of Electronic Imaging
& the Visual Art (EVA’06 Vienna) - Digital Cultural Heritage Essential for Tourism, pp. 83–90.
Leute, U., 1987. Archaeometry: An Introduction to Physical
Methods in Archaeology and the History of Art. John Wiley &
Sons.
Pottmann, H., Steiner, T., Hofer, M., Haider, C. and Hanbury, A.,
2004. The isophotic metric and its applications to feature sensitive morphology on surfaces. In: Computer Vision - ECCV
2004, Part IV, volume 3024 of Lecture Notes in Computer Science, Springer, pp. 560–572.
Reindel, M. and Cuadrado, J., 2001. Los Molinos und La Muna.
Zwei Siedlungszentren der Nasca-Kultur in Palpa, Südperu. In:
P. von Zabern (ed.), In Beiträge zur Allgemeinen und Vergleichenden Archäologie, Vol. 21.
Reisert, M., Ronneberger, O. and Burkhardt, H., 2006. An efficient gradient based registration technique for coin recognition.
In: Proc. of the Muscle CIS Coin Competition Workshop, Berlin,
Germany, pp. 19–31.
Rössl, C., Kobbelt, L. and Seidel, H., 2000. Extraction of feature lines on triangulated surfaces using morphological operators.
In: Proc. Smart Graphics 2000, AAAI Spring Symposium, AAAI
Press.
Sablatnig, R., Menard, C. and Dintsis, P., 1991. A Preliminary
Study on Methods for a Pictorial Acquisition of Archaeological Finds. Technical Report PRIP-TR-010, Vienna University
of Technology, Institute of Computer Aided Automation, Pattern
Recognition and Image Processing Group.
Stern, E., 1994, extended 2000. Dor - Ruler of the Seas. Israel
Exploration Society.
Strubecker, K., 1969.
Differentialgeometrie: Theorie der
Flächenmetrik. Sammlung Göschen, Vol. 1197, De Gruyter,
Berlin.
Sun, Y. and Abidi, M., 2001. Surface Matching by 3D Point’s
Fingerprint. In: Proc.. Eighth IEEE International Conference on
Computer Vision (ICCV’01), Vol. 2, pp. 263–269.
Thodberg, H., 2003. Minimum Description Length Shape and
Appearance Models. Vol. 2732.
Tosovic, S., 2002. Adaptive 3D Modeling of Objects by combining Shape from Silhouette and Shape from Structured Light.
Master’s thesis, Vienna University of Technology, Vienna University of Technology, Institute of Computer Aided Automation,
Pattern Recognition and Image Processing Group.
Utili, F. and Dolmazon, P., 2002. Bemerkungen zu drei attischen
Trinkgefäßen im Reiss-Museum in Mannheim. In: M. Bentz
(ed.), Vasenforschung und Corpus vasorum antiquorum: Standortbestimmung und Perspektiven, Beihefte zum CVA 1, p. 37.
van der Maaten, L. J. and Poon, P., 2006. Coin-o-matic: A fast
system for reliable coin classification. In: Proc. of the Muscle
CIS Coin Competition Workshop, Berlin, Germany, pp. 07–18.
van der Maaten, L. J. and Postma, E. O., 2006. Towards automatic
coin classification. In: Proc. of the EVA-Vienna 2006, Vienna,
Austria, pp. 19–26.
Wieczorek, A. and Tellenbach, M., 2002. Exkurs zur Frage der
Drehscheibenkeramik. In: An die Mächte der Natur - Mythen
der altperuanischen Nasca-Indianer. Katalog zur Ausstellung im
Reiss-Engelhorn-Museum. Mainz: Philipp von Zabern., pp. 54–
63.
Willis, R., 2004. Stochastic 3D Geometric Models for Classification, Deformation, and Estimation. Ph.D. thesis. Brown University, Rhode Island, USA.